Mistral Patches Six High-Severity Shell Vulnerabilities in Vibe CLI

Mistral AI released version 2.25.4 of the Mistral Vibe CLI on September 12, 2026, to address six high-severity vulnerabilities in its shell permission checks. [1] These vulnerabilities, tracked from CVE-2026-87983 through CVE-2026-87988, allowed specially crafted shell commands to execute without the expected user approval prompts. [1] The flaws occurred because the system's permission checks failed to account for all command syntax, file paths, or options. [1] Malicious instructions embedded in content processed by the agent could trigger these unapproved commands. [1]

Depending on the specific vulnerability, unapproved commands could execute code using the permissions of the user running Vibe. [1] Attackers could also read, create, or overwrite files located outside the designated workspace, which could expose accessible credentials and source code. [1] The overall impact depends heavily on the system permissions, files, and credentials available to the Vibe process. [1] While Mistral implemented partial fixes for CVE-2026-87988 in versions 2.20.0 and 2.23.0, version 2.25.4 resolves the remaining reported command-option bypasses for both legacy and managed shell implementations. [1]

The release of Mistral Vibe 2.25.4 highlights the ongoing challenge of securing AI agents that interact with system shells. Because malicious instructions in agent-processed content could trigger these bypasses, organizations using the CLI must recognize that prompt injection or poisoned data can directly lead to unauthorized code execution and credential theft. The phased remediation of CVE-2026-87988, which required partial fixes in earlier versions following HiddenLayer's advisories before full resolution in 2.25.4, demonstrates the complexity of parsing diverse shell command syntax and path variables securely. What the sources don't address: whether Mistral observed any of these vulnerabilities being actively exploited in the wild prior to the patch release.